src/event_dispatcher.cpp
| Line | Branch | Exec | Source |
|---|---|---|---|
| 1 | /** | ||
| 2 | * @file event_dispatcher.cpp | ||
| 3 | * @brief The dispatcher's per-thread emit chain and rundown drain. | ||
| 4 | * | ||
| 5 | * EventDispatcher is otherwise a header-only template. What lives here is the part that cannot: the emit chain is | ||
| 6 | * backed by a Win32 TLS index, and the installed detail/ header must stay Win32-free. | ||
| 7 | */ | ||
| 8 | |||
| 9 | #include "DetourModKit/detail/event_dispatcher.hpp" | ||
| 10 | |||
| 11 | #include "internal/drain_backoff.hpp" | ||
| 12 | #include "platform.hpp" | ||
| 13 | |||
| 14 | #include <atomic> | ||
| 15 | #include <thread> | ||
| 16 | |||
| 17 | namespace DetourModKit::detail | ||
| 18 | { | ||
| 19 | namespace | ||
| 20 | { | ||
| 21 | /** | ||
| 22 | * @brief The TLS index holding this thread's EmitFrame chain head. | ||
| 23 | * @details Reserved by subscribe() before any emit of a subscribed dispatcher can run, so a dispatcher with | ||
| 24 | * subscribers always has a valid index. TLS_OUT_OF_INDEXES means the process had none to give, which | ||
| 25 | * leaves every emit untracked and every rundown Unwaitable rather than wrong. | ||
| 26 | */ | ||
| 27 | std::atomic<DWORD> s_emit_tls_index{TLS_OUT_OF_INDEXES}; | ||
| 28 | |||
| 29 | std::atomic<std::uint32_t> s_untracked_emit_frames{0}; | ||
| 30 | } // namespace | ||
| 31 | |||
| 32 | 10338 | bool ensure_emit_frame_tls() noexcept | |
| 33 | { | ||
| 34 |
2/2✓ Branch 9 → 10 taken 10241 times.
✓ Branch 9 → 11 taken 97 times.
|
10338 | if (s_emit_tls_index.load(std::memory_order_acquire) != TLS_OUT_OF_INDEXES) |
| 35 | { | ||
| 36 | 10241 | return true; | |
| 37 | } | ||
| 38 | |||
| 39 | // Reserve first and publish with a CAS rather than serializing on a lock. A mutex here would be a static with a | ||
| 40 | // destructor on a path a late subscriber can still reach after this TU's statics are torn down; racing losers | ||
| 41 | // simply return their surplus index instead. | ||
| 42 | 97 | const DWORD index = ::TlsAlloc(); | |
| 43 |
1/2✗ Branch 12 → 13 not taken.
✓ Branch 12 → 14 taken 97 times.
|
97 | if (index == TLS_OUT_OF_INDEXES) |
| 44 | { | ||
| 45 | ✗ | return false; | |
| 46 | } | ||
| 47 | 97 | DWORD unreserved = TLS_OUT_OF_INDEXES; | |
| 48 | 97 | if (!s_emit_tls_index | |
| 49 |
1/2✗ Branch 19 → 20 not taken.
✓ Branch 19 → 21 taken 97 times.
|
97 | .compare_exchange_strong(unreserved, index, std::memory_order_release, std::memory_order_relaxed)) |
| 50 | { | ||
| 51 | ✗ | ::TlsFree(index); | |
| 52 | } | ||
| 53 | 97 | return true; | |
| 54 | } | ||
| 55 | |||
| 56 | 1741645 | bool push_emit_frame(EmitFrame &frame) noexcept | |
| 57 | { | ||
| 58 | 1736589 | const DWORD index = s_emit_tls_index.load(std::memory_order_acquire); | |
| 59 |
1/2✗ Branch 9 → 10 not taken.
✓ Branch 9 → 11 taken 1736589 times.
|
1736589 | if (index == TLS_OUT_OF_INDEXES) |
| 60 | { | ||
| 61 | ✗ | return false; | |
| 62 | } | ||
| 63 | 1736589 | frame.prev = static_cast<EmitFrame *>(::TlsGetValue(index)); | |
| 64 | // A store can still fail: an index past the TEB's inline slots is backed by a lazily heap-allocated | ||
| 65 | // expansion array. Report it rather than leave the chain claiming this thread is elsewhere. | ||
| 66 | 1734109 | return ::TlsSetValue(index, &frame) != FALSE; | |
| 67 | } | ||
| 68 | |||
| 69 | 1606524 | void pop_emit_frame(const EmitFrame &frame) noexcept | |
| 70 | { | ||
| 71 | 1594949 | const DWORD index = s_emit_tls_index.load(std::memory_order_acquire); | |
| 72 |
1/2✗ Branch 9 → 10 not taken.
✓ Branch 9 → 11 taken 1594949 times.
|
1594949 | if (index == TLS_OUT_OF_INDEXES) |
| 73 | { | ||
| 74 | ✗ | return; | |
| 75 | } | ||
| 76 | // The matching push succeeded, so the expansion array for this index already exists on this thread and this | ||
| 77 | // store cannot fail for want of one. | ||
| 78 | 1594949 | (void)::TlsSetValue(index, frame.prev); | |
| 79 | } | ||
| 80 | |||
| 81 | 7 | bool thread_is_emitting_dispatcher(const void *dispatcher) noexcept | |
| 82 | { | ||
| 83 | 7 | const DWORD index = s_emit_tls_index.load(std::memory_order_acquire); | |
| 84 |
2/2✓ Branch 9 → 10 taken 1 time.
✓ Branch 9 → 11 taken 6 times.
|
7 | if (index == TLS_OUT_OF_INDEXES) |
| 85 | { | ||
| 86 | 1 | return false; | |
| 87 | } | ||
| 88 |
2/2✓ Branch 16 → 13 taken 2 times.
✓ Branch 16 → 17 taken 5 times.
|
7 | for (const auto *node = static_cast<const EmitFrame *>(::TlsGetValue(index)); node != nullptr; |
| 89 | 1 | node = node->prev) | |
| 90 | { | ||
| 91 |
2/2✓ Branch 13 → 14 taken 1 time.
✓ Branch 13 → 15 taken 1 time.
|
2 | if (node->dispatcher == dispatcher) |
| 92 | { | ||
| 93 | 1 | return true; | |
| 94 | } | ||
| 95 | } | ||
| 96 | 5 | return false; | |
| 97 | } | ||
| 98 | |||
| 99 | 10342 | bool thread_is_emitting_type(const void *type_tag) noexcept | |
| 100 | { | ||
| 101 | 10342 | const DWORD index = s_emit_tls_index.load(std::memory_order_acquire); | |
| 102 |
2/2✓ Branch 9 → 10 taken 98 times.
✓ Branch 9 → 11 taken 10244 times.
|
10342 | if (index == TLS_OUT_OF_INDEXES) |
| 103 | { | ||
| 104 | 98 | return false; | |
| 105 | } | ||
| 106 |
2/2✓ Branch 16 → 13 taken 3 times.
✓ Branch 16 → 17 taken 10241 times.
|
10244 | for (const auto *node = static_cast<const EmitFrame *>(::TlsGetValue(index)); node != nullptr; |
| 107 | ✗ | node = node->prev) | |
| 108 | { | ||
| 109 |
1/2✓ Branch 13 → 14 taken 3 times.
✗ Branch 13 → 15 not taken.
|
3 | if (node->type_tag == type_tag) |
| 110 | { | ||
| 111 | 3 | return true; | |
| 112 | } | ||
| 113 | } | ||
| 114 | 10241 | return false; | |
| 115 | } | ||
| 116 | |||
| 117 | 10341 | std::atomic<std::uint32_t> &untracked_emit_frames() noexcept | |
| 118 | { | ||
| 119 | 10341 | return s_untracked_emit_frames; | |
| 120 | } | ||
| 121 | |||
| 122 | 7 | Rundown drain_gate(EntryGate &gate, const void *dispatcher) noexcept | |
| 123 | { | ||
| 124 |
5/6✓ Branch 3 → 4 taken 6 times.
✓ Branch 3 → 12 taken 1 time.
✗ Branch 11 → 12 not taken.
✓ Branch 11 → 13 taken 6 times.
✓ Branch 14 → 15 taken 1 time.
✓ Branch 14 → 16 taken 6 times.
|
13 | if (thread_is_emitting_dispatcher(dispatcher) || s_untracked_emit_frames.load(std::memory_order_seq_cst) != 0) |
| 125 | { | ||
| 126 | 1 | return Rundown::Unwaitable; | |
| 127 | } | ||
| 128 | |||
| 129 | // The tombstone is already published (seq_cst) by the caller, and every invocation increments in_flight | ||
| 130 | // seq_cst before rechecking it. Those two orders are the Dekker seam: an entrant that misses the tombstone | ||
| 131 | // is guaranteed visible here, so reaching zero means no invocation remains and none can start. | ||
| 132 | // | ||
| 133 | // This wait has no timeout. The tombstone closes the entrant set, so it cannot grow. A tracked handler can | ||
| 134 | // still remain parked indefinitely. Self-entry deadlocks, and an untracked entry makes completion | ||
| 135 | // unprovable, so both cases are refused above. Backoff limits CPU use while a tracked handler finishes. | ||
| 136 | 6 | DrainBackoff backoff; | |
| 137 |
2/2✓ Branch 25 → 17 taken 132 times.
✓ Branch 25 → 26 taken 6 times.
|
276 | while (gate.in_flight.load(std::memory_order_seq_cst) != 0) |
| 138 | { | ||
| 139 | 132 | backoff.pause(); | |
| 140 | } | ||
| 141 | 6 | return Rundown::Drained; | |
| 142 | } | ||
| 143 | } // namespace DetourModKit::detail | ||
| 144 |